Start with your problem
Read one useful piece now. Follow a path when you need the whole system.
Build an agent
Choose the execution loop, expose useful tools, and bound what can change.
- Before You Build Another Agent
Start with one capable generalist. Add delegation only when the task earns it.
- Implementing Progressive Disclosure for Agent Tools
Move from a tool catalogue to discoverable, executable interfaces.
- Trustworthy Autonomy
Design evidence, ownership, permissions, and recovery as runtime guarantees.
Evaluate a system
Define success, score the right things, and test whether the result holds.
- How to Write and Score an Agent Eval
A worked incident, scoring explorer, and runnable Python companion.
- Checkpoint Replay for Long-Horizon Agent Evaluation
Freeze what the agent could know at each decision.
- Classifiers: When Labels Become Decisions
Build a label policy and measure the errors your decision can tolerate.
Improve what works
Turn a failure into a bounded experiment and preserve what survives review.
- The Quality Flywheel
Move from failure evidence to experiments, release decisions, and reusable lessons.
- Building an Agent That Improves Its Own Harness
Search over the harness while keeping promotion outside the optimizer.
- How Intelligence Finds Its Way
Understand the environments, examples, and reversible actions that make exploration useful.
A small working glossary
- Harness
- The execution loop, tools, state, policies, and feedback surrounding a model. Read more
- Progressive disclosure
- Reveal an interface or evidence when it becomes relevant, while keeping the route to it discoverable. Read more
- Checkpoint replay
- Resume a recorded decision with only the information available at that checkpoint. Read more
- Provenance
- The source, transformations, scope, and time that make an evidence claim inspectable. Read more
- Ground truth
- A versioned reference judgment under a defined policy, including disagreement and unresolved cases. Read more
- Calibration
- Whether stated probabilities match observed frequencies over comparable cases. Read more
- Selective risk
- The error rate among accepted decisions as uncertain cases are deferred. Read more
- Authority gate
- A separate release or runtime condition for permitted actions; correctness cannot compensate for violating it. Read more
- Compare and swap (CAS)
- Apply a write only if the stored version still matches the version the worker read. Read more
- Paired evaluation
- Compare candidates on the same tasks and account for which tasks changed outcome. Read more
- Counterfactual
- What would have happened under a different action; a fixed recording alone cannot establish it. Read more
- Negative caching
- Cache an absence or failure under an explicit validity policy; stable absence differs from a transient outage. Read more